Skip to content

CHANGELOG.md

Latest commit

 

History

History
146 lines (109 loc) · 8.48 KB

CHANGELOG.md

File metadata and controls

146 lines (109 loc) · 8.48 KB

Changelog

All notable changes to Project Aura are documented here.

Format: Keep a Changelog — Added, Changed, Fixed, Removed Versioning: Semantic Versioning

Unreleased

Changes on feature branches not yet merged to main.

Added

  • docs/USE_CASES.md — 16 procurement use cases: 3 core demo scenarios, 8 SSA contract type scenarios, and 5 extended real-world scenarios (volume discounts, mandate splits, geo-restrictions, payment thresholds, audit trails)

In Progress

  • test/architect-e2e — E2E integration test for the full Scout → Sentinel → Closer pipeline (#2)
  • feat/kagent-manifest — Kubernetes Agent CRD review and kagent apply --dry-run validation (#1)

[feat/policy-engine] — Unreleased

Policy Engine feature branch (#15). Adds three-stage governance rules for procurement.

Added

Policy Engine (tools/)

  • tools/policy_store.py — PolicyStore singleton with 6 default rules (SPENDING_LIMIT, GEO_RESTRICTION, CATEGORY_ALLOWLIST, APPROVAL_THRESHOLD, CERTIFICATION_REQUIRED, RATE_LIMIT); atomic JSON persistence to tmp/policies.json; ReviewStore for human-in-the-loop payment review queue
  • tools/policy_tools.py — Three evaluation functions consumed by agents:
    • evaluate_procurement_policy(request) — pre-flight gate (category, spending, rate limits)
    • evaluate_vendor_policy(vendor, amount) — vendor gate (geo-restriction, certifications, approval thresholds)
    • evaluate_payment_policy(mandate, user_id) — payment gate (approval thresholds, daily spend limits)
    • RateLimitStore — sliding-window per-user rate limiting (in-memory)
    • DailySpendStore — per-user daily spend accumulation (in-memory)

Governor Agent (agents/)

  • agents/governor.py — New pre-flight LlmAgent; inserted first in the pipeline before Scout; outputs POLICY_CLEAR / POLICY_WARNINGS / POLICY_REVIEW_REQUIRED / POLICY_BLOCKED

Pipeline Changes

  • agents/architect.py — Updated pipeline: Governor → Scout → Sentinel → Closer (previously Scout → Sentinel → Closer)
  • agents/sentinel.py — Added evaluate_vendor_policy tool; now runs both BMS compliance check AND vendor policy check per vendor
  • agents/closer.py — Added evaluate_payment_policy tool; checks governor_results and sentinel_results for policy/compliance blocks before any payment settlement

Policy Management REST API (main.py)

  • GET /policies — list all policy rules
  • POST /policies — create rule (requires X-Admin-Token header)
  • GET /policies/{rule_id} — get single rule
  • PUT /policies/{rule_id} — partial update (requires X-Admin-Token)
  • DELETE /policies/{rule_id} — delete rule, returns 204 (requires X-Admin-Token)
  • GET /reviews — list pending payment reviews
  • POST /reviews/{id}/approve — approve a review (requires X-Admin-Token)
  • POST /reviews/{id}/reject — reject a review (requires X-Admin-Token)

Kubernetes (kagent.yaml)

  • aura-admin-token Kubernetes Secret manifest for AURA_ADMIN_TOKEN
  • Governor Agent CRD with evaluate_procurement_policy MCP tool
  • All agents (Scout, Sentinel, Closer, Architect) updated with AURA_ADMIN_TOKEN env var
  • Sentinel and Closer updated with new policy tool registrations

Tests (tests/)

  • tests/test_policy_tools.py — 27 new unit tests covering all 6 rule types, rate limiting, daily spend, snapshot hashing, and disabled-rule bypass

Documentation

  • docs/AGENT_FLOW.md — Rewritten with three Mermaid sequence diagrams: Happy Path, Policy Block (Governor halts pre-flight), Compliance Block (Sentinel → Closer aborts)

0.1.0-hackathon — 2026-03-11

Hackathon prototype delivered at Google AI Agent Labs Oslo 2026 — Team 6. All core agent logic, tooling, infrastructure, and documentation complete.

Added

Agents (agents/)

  • architect.py — Root LlmAgent orchestrator; parses procurement intent, chains Scout → Sentinel → Closer via SequentialAgent
  • scout.py — UCP vendor discovery agent; calls discover_vendors(), flags risky vendors (country XX, suspicious pricing), writes scout_results to session state
  • sentinel.py — KYC/AML compliance gate; runs every vendor through verify_vendor_compliance(), outputs COMPLIANCE_BLOCKED on any rejection
  • closer.py — AP2 payment settlement agent; generates IntentMandate (W3C VC), settles via settle_cart_mandate(), aborts if Sentinel blocked

Tools (tools/)

  • ucp_tools.py — discover_vendors(query) — mock UCP vendor database of 4 vendors (including deliberately blacklisted ShadowHardware), sorted by unit price
  • compliance_tools.py — verify_vendor_compliance(vendor_name) — BMS KYC/AML mock; deterministic 64-char ComplianceHash for approved vendors; AML_BLACKLIST rejection for ShadowHardware
  • ap2_tools.py — generate_intent_mandate() + settle_cart_mandate() — AP2 mock with W3C VC IntentMandate structure, ECDSA-P256 proof, 5000 USD cap, AP2-* settlement IDs

API Server

  • main.py — FastAPI server (port 8080) with Google ADK Runner; POST /run (sync) + POST /run/stream (SSE); GET /health liveness probe; InMemorySessionService for session state

Infrastructure

  • Dockerfile — Multi-stage Python 3.12 build; non-root aura user; healthcheck on /health
  • kagent.yaml — kagent.dev/v1alpha2 manifests for all 4 agents + shared ModelConfig (Vertex AI Gemini 2.5 Flash, us-central1) + ClusterIP Service
  • .env / .env.example — GCP project config (ai-agent-labs-oslo-26-team-6, us-central1)

CI/CD (.github/workflows/)

  • ci.yml — ruff lint → pytest → Docker build (runtime stage); triggers on every push and PR to main
  • cd.yml — GCP Artifact Registry push + kagent apply; triggers on push to main; workflow_dispatch with skip_deploy safety input

Tests (tests/)

  • test_ucp_tools.py — 9 tests: vendor count, sort order, ShadowHardware XX flag, required fields, query-agnostic behaviour
  • test_compliance_tools.py — 17 tests: blacklist (exact + case variants + whitespace), 64-char hex hash, determinism within hour, unique hashes per vendor
  • test_ap2_tools.py — 27 tests: mandate structure, UUID ID, amount cap, ECDSA proof, settlement ID format, missing-hash guard, wrong-type guard
  • conftest.py — pytest asyncio mode + integration marker
  • Total: 91 tests, 0 failures

Documentation (docs/)

  • ARCHITECTURE.md — System diagram (Mermaid), component table, data flow, technology stack
  • AGENT_FLOW.md — Sequence diagrams: happy path + compliance-blocked path
  • DATA_MODEL.md — Class diagram + JSON examples for all 4 data structures
  • DEPLOYMENT.md — Kagent install, GCP setup, Docker build/push, verification commands, cloud-agnostic model swap guide
  • PROTOCOLS.md — UCP, AP2, BMS protocol explanations; mock-vs-real integration paths
  • DEMO_SCRIPT.md — Full pitch script with timing, live demo steps, judge Q&A
  • docs/README.md — Index of all documentation files

Project Hygiene

  • README.md — Agent squad table, quick-start guide, Mermaid flowchart, docs index
  • pyproject.toml — pytest config (asyncio_mode=auto, testpaths, --tb=short -v)
  • requirements.txt — google-adk, google-cloud-aiplatform, fastapi, uvicorn, cryptography, pydantic, pytest, pytest-asyncio, httpx
  • LICENSE — Apache 2.0
  • .gitignore — Python, venv, IDE, Zone.Identifier metadata exclusions
  • .github/pull_request_template.md — Summary, type checkboxes, compliance checklist
  • dashboard/ — Streamlit presentation dashboard for live demo

GitHub Project Setup

  • 6 labels: agent, infra, testing, ci-cd, compliance, docs
  • 2 milestones: Demo Day — Mar 11 2026, v0.1 Post-Hackathon
  • 8 issues (#1–#8) covering all components, with labels, milestones, and linked branches
  • 8 feature branches: feat/kagent-manifest, test/architect-e2e, test/scout-unit, test/sentinel-unit, test/closer-unit, feat/ci-pipeline, feat/cd-pipeline, chore/pr-template

0.0.1 — 2026-03-11

Initial scaffolding commit.

Added

  • Project structure: agents/, tools/, tests/, docs/
  • __init__.py stubs for all packages
  • Initial requirements.txt
⚡